Newest at the top
2024-05-16 09:36:41 +0200 | bilegeek | (~bilegeek@2600:1008:b010:1cfe:3776:17b8:1dc2:3fdf) (Quit: Leaving) |
2024-05-16 09:33:31 +0200 | <tomsmeding> | but true |
2024-05-16 09:33:02 +0200 | <tomsmeding> | with slightly more overhead |
2024-05-16 09:32:35 +0200 | <c_wraith> | and a TVar-based approach lets you write much more obviously-correct code that works under the write-light conditions. |
2024-05-16 09:30:55 +0200 | bo_ | (~bo@198.red-83-56-252.dynamicip.rima-tde.net) (Ping timeout: 256 seconds) |
2024-05-16 09:30:46 +0200 | <tomsmeding> | hence the "write-light" observed above |
2024-05-16 09:30:38 +0200 | <tomsmeding> | of course |
2024-05-16 09:30:31 +0200 | <c_wraith> | in practice, an MVar-based implementation will perform way better with heavy writes. |
2024-05-16 09:29:30 +0200 | <tomsmeding> | and observing that this cleverer implementation does not have stronger preconditions than the MVar-based one |
2024-05-16 09:29:28 +0200 | kuribas` | (~user@ip-188-118-57-242.reverse.destiny.be) |
2024-05-16 09:29:12 +0200 | <tomsmeding> | I'm looking at an idealised implementation in terms of `MVar (Map Key Value)` |
2024-05-16 09:28:37 +0200 | <c_wraith> | You're adding extra semantics. I'm looking at the type. :P |
2024-05-16 09:28:09 +0200 | <tomsmeding> | I that would be a requirement from the interface of getRefForKey anyway, regardless of the implementation |
2024-05-16 09:27:45 +0200 | _bo | (~bo@198.red-83-56-252.dynamicip.rima-tde.net) |
2024-05-16 09:27:07 +0200 | <c_wraith> | Ok, it works on the assumption that the value you want to insert doesn't depend on whether there was already a key there or not (or depends trivially, with some sort of commutative operation) |
2024-05-16 09:23:57 +0200 | frumon | (~Frumon@user/Frumon) |
2024-05-16 09:22:25 +0200 | <tomsmeding> | write-light, then |
2024-05-16 09:22:17 +0200 | <tomsmeding> | :D |
2024-05-16 09:21:41 +0200 | <Axman6> | actually that's not true, since things were only read when a uses used the webservice, which no one ever did =) |
2024-05-16 09:21:39 +0200 | sawilagar | (~sawilagar@user/sawilagar) |
2024-05-16 09:21:05 +0200 | <Axman6> | yeah - which my one use case for this particular design was =) |
2024-05-16 09:20:44 +0200 | <tomsmeding> | also needs a read-heavy workload |
2024-05-16 09:20:31 +0200 | <tomsmeding> | but if that's enough, it's nice |
2024-05-16 09:20:27 +0200 | <tomsmeding> | gets you a limited interface though, to the map |
2024-05-16 09:20:20 +0200 | <tomsmeding> | neat |
2024-05-16 09:20:00 +0200 | <Axman6> | but in the case where the key exists no synchronisation is needed |
2024-05-16 09:19:49 +0200 | <Axman6> | of the threads* |
2024-05-16 09:19:46 +0200 | <tomsmeding> | interesting |
2024-05-16 09:19:21 +0200 | <Axman6> | the code above should handle that fine, since it uses atomicModifyIORef in the update case, only one of the can put their key in |
2024-05-16 09:18:42 +0200 | <tomsmeding> | time-of-check to time-of-use race condition |
2024-05-16 09:18:30 +0200 | <tomsmeding> | oh hm fair |
2024-05-16 09:18:10 +0200 | <c_wraith> | so do I |
2024-05-16 09:18:08 +0200 | <Axman6> | yes |
2024-05-16 09:18:03 +0200 | <tomsmeding> | judging from the getRefForKey code |
2024-05-16 09:17:53 +0200 | <tomsmeding> | c_wraith: I think Axman6 meant `IORef (Map Key (IORef Value))` |
2024-05-16 09:17:43 +0200 | <c_wraith> | It also assumes you never have two threads trying to insert the same key at the same time. |
2024-05-16 09:17:14 +0200 | <Axman6> | this also assumes you never delete keys (use Maybe (IORef Value)) then) |
2024-05-16 09:16:46 +0200 | <Axman6> | oh and the Just case for the first case uses the value returned |
2024-05-16 09:16:28 +0200 | <Axman6> | (missing `old` in the Map.insert but that should work) |
2024-05-16 09:15:37 +0200 | <c_wraith> | I'm not even sure it can be used safely. If you look at the Map and determine you need to insert a new key/value pair, you can't do that from within an atomicModifyIORef |
2024-05-16 09:15:20 +0200 | <Axman6> | getRefForKey outer key = do exists <- Map.lookup key <$> readIORef ref; case exists of Nothing -> do newRef <- newIORef; atomicModifyIORef ref (\old -> case Map.lookup key old of Nothing -> (Map.insert key newRef, newRef); Just race -> (old, race)) |
2024-05-16 09:13:09 +0200 | <Axman6> | it can be used safely, you just wrap things in a safe interface |
2024-05-16 09:11:56 +0200 | emmanuelux | (~emmanuelu@user/emmanuelux) (Quit: au revoir) |
2024-05-16 09:10:14 +0200 | <c_wraith> | It's begging to introduce race conditions |
2024-05-16 09:09:40 +0200 | <c_wraith> | that last one is dangerous |
2024-05-16 09:08:57 +0200 | kuribas` | (~user@2a02:1808:8:dd22:98a8:ccd5:a4c3:db95) (Ping timeout: 256 seconds) |
2024-05-16 09:08:08 +0200 | <Axman6> | yeah definitely, and can also lead to less contention on shared values - the difference between n IORef (Map Key Value) and Map Key (IORef value) (or even IORef (Map Key (IORef Value)) might make significant performance differences. |
2024-05-16 09:06:18 +0200 | euleritian | (~euleritia@ip4d16fc38.dynamic.kabel-deutschland.de) |
2024-05-16 09:06:00 +0200 | euleritian | (~euleritia@dynamic-176-006-196-153.176.6.pool.telefonica.de) (Read error: Connection reset by peer) |
2024-05-16 09:03:05 +0200 | <tomsmeding> | having transactions sometimes relieves you of a whole lot of trouble in designing your data structures and methods |